Skip to content

100 Prompts to Test Your Brand’s AI Visibility in Arab Markets

100 Prompts to Test Your Brand's AI Visibility in Arab Markets
  • Reading time 23 min
  • Summarize with AI
  • Nov. 6, 2026

These 100 prompts test whether ChatGPT and Google AI Mode name your brand. Half are English and half Arabic, written in Gulf dialect, Egyptian dialect and formal Arabic across nine buyer intents, with a free scoring sheet. We ran 20 of them on 19 October 2026, and every answer came back in the prompt’s language. ChatGPT showed sponsored units in 13 of 20 answers at night, none of 20 in an afternoon repeat; AI Mode showed none in either run.

One run is one draw, so we asked both engines the same 20 prompts again, ChatGPT about 14 hours later and Google AI Mode about 60 hours later. A SparkToro and Gumshoe study of 2,961 runs in late 2025 found ChatGPT and Google’s AI returned the same brand list less than once in 100 repeats. In the repeats, ChatGPT kept its first pick on 9 of the 14 prompts with recommendations and AI Mode on 7 of 14. Each engine’s two runs shared 0.41 and 0.42 of their recommended brands on average, against 0.25 between engines: each differed from the other more than from itself.

100
prompts: 50 English, 17 Gulf Arabic, 17 Egyptian Arabic and 16 formal Arabic, across nine intents
VOCTOS prompt library, October 2026
13 vs 0
of 20 signed-out ChatGPT answers showed a sponsored unit, at night against the afternoon repeat; Google AI Mode showed none in either run
VOCTOS test, 19 and 21 Oct 2026, runs 1 and 2
0 of 4
Gulf-dialect prompts that Google AI Mode answered in Gulf Arabic; ChatGPT matched 3 of 4
VOCTOS test, 19 Oct 2026, run 1
<1%
of repeated runs of the same prompt returned the same brand list
SparkToro and Gumshoe, 2,961 runs, late 2025

Get the files: download all 100 AI visibility test prompts as an Excel workbook, which adds How to use, Score log and Summary tabs. The same library as a CSV file opens in any spreadsheet, and you can check every Arabic prompt against its English meaning.

What a prompt test measures, and what it cannot

A prompt test measures presence: whether an AI engine names, recommends or cites your brand when a buyer asks a real question today. It cannot measure readiness, the site and entity signals that make you citable, and a single answer proves little, because engines rarely return the same list twice.

Presence and readiness need different tools. A prompt test shows where you stand in this week’s answers. The crawler access, structure and entity work behind those answers is readiness, which our AI visibility score measures on its own.

Answers also move between runs. Ahrefs tracked 43,000 keywords for a month and saw AI Overview content change 70% of the time between observations, in a study published in November 2025. So read every result you get as a share of many answers, never as a rank.

A prompt test also stops at the answer and cannot tell you who clicked. Even the clicks that arrive can be misfiled, as when ChatGPT visits landed in GA4 as Direct traffic in October 2026.

How many prompts do you need?

For one market and one language, tool makers’ guidance starts between 20 and 100 prompts. If you sell in Arab markets, count each language separately, because English and Arabic answers share few sources. Start with 20 prompts for your first test, track about 50 a month, and keep all 100 for a one-off audit.

Source Starting set Branded prompts Language and location
Profound (guide, Feb 2026) 100 to start; its users track 100 to 1,000 Mostly unbranded; competitor and alternatives prompts a small share Phrase prompts the way the target market speaks; region-based prompting in 80+ countries
Peec AI (documentation) No fixed number; its prompt discovery sets the mix 20% branded by default, with a 25/50/25 informational, commercial and transactional split One country per prompt; 200 characters at most
Otterly.AI (onboarding, Sep 2026) About 100 per country, per language, per product category; 50 at minimum Unbranded, middle to bottom of the funnel Tag every prompt by country and language; 2 categories in 2 countries and 2 languages come to about 400
Semrush (AI Visibility toolkit) 25 included at USD 99 a month; 50 more cost USD 60 a month Four prompt types: revenue, reputation, competitor and gap Tracking in 220+ countries and territories; Saudi Arabia and the UAE are research regions, Egypt is not listed
Ahrefs (Custom Prompts) 5, 10 or 20 by plan; from 83 on Enterprise Tag by topic, funnel stage or campaign One check per prompt, platform, location and update
SE Ranking (guide, Apr 2026) 20 to 40 in total 5 to 10 brand prompts, tracked apart from category prompts Narrow broad prompts with geographic qualifiers; run on 2 to 3 engines
HubSpot Germany (guide in German, 2026) 25 to 50; its AEO tool takes up to 50 Brand names in prompts produce false positives “As long as necessary, as short as possible”, since extra context widens the field

Read those counts as commercial defaults as much as research. Semrush’s 25 is what its base plan includes, and HubSpot’s 50 is its tool’s limit. Before you compare vendors, check how the paid tools measure the same thing in three different ways.

Arab markets add one rule: count Arabic and English as two sets. Otterly already counts per language, and in our Saudi and Egyptian studies, switching language changed sources and brands more than a plain repeat did. Keep branded prompts to about a quarter of your set (Peec’s default is 20%) and report them apart. A brand named in the prompt is nearly always named in the answer.

Arab markets add one rule: count Arabic and English as two sets.

How the 100 prompts are organised

The library crosses nine buyer intents with four funnel stages and four language arms: 50 English templates, 17 Gulf Arabic, 17 Egyptian Arabic and 16 formal Arabic. Most Arabic rows are twins of an English row, so you compare like with like, and square-bracket placeholders let you fill the sheet with one find-and-replace list.

Intent English Formal Arabic Gulf Egyptian Two English templates from the library
Category discovery 7 3 2 2 “What are the best [category] in [city]?” and the search-box version “best [category] [city]”
Comparison 6 2 2 2 “[brand] vs [competitor]: which is better for [use case] in [country]?” and “Compare the top 3 [category] in [city] on price, quality and waiting time.”
Problem-first 6 1 2 2 “I have [problem]. Where can I find a [category] open right now in [city]?” and “How do I fix [problem]? Which tools or companies should I use?”
Price 6 2 2 2 “How much does [service] cost in [city]?” and “Which online stores in [country] let me buy [product] in installments with no interest?”
Trust and reviews 6 2 2 2 “Which [category] in [city] have the best reviews from real customers?” and “Is [brand] legit? What do customers say about it?”
Local 5 1 2 2 “Which [category] in [area] have a female [professional]?” and “Which [category] in [city] accept [insurance provider] insurance?”
Brand facts 6 2 2 2 “Is [brand] licensed by [regulator] in [country]?” and “What is [brand]’s official customer service number in [country]?”
Alternatives 4 1 1 2 “What are the best alternatives to [competitor] in [country]?” and “I’m unhappy with [competitor]. Who else should I try in [city]?”
Buyer stage 4 2 2 1 “Where can I buy an original [product] online in [country] with an official local warranty?” and “Which [category] in [city] let me book online and pay in cash?”
Total 50 16 17 17 All 100 rows, with notes, are in the workbook

Thirty-nine prompts name a brand or a competitor, 20 in English and 19 in Arabic, because brand facts, comparisons and alternatives need a name. The notes column flags each row as branded, competitor-branded or unbranded. By funnel stage the split is 10 awareness, 55 consideration, 30 decision and 5 post-purchase, because informational questions rarely bring up brands.

The rows also carry buying habits that foreign templates miss. Prompts cover installments through Tabby, Tamara or valU, cash on delivery, agent warranties and counterfeit stock, plus female practitioners, insurance networks, Ramadan hours and White Friday. Five are typed like a search box, one of them in Arabic without the hamza, and two experimental rows use Arabizi, Arabic in Latin letters.

Why test in Arabic and English at all?

If your buyers search in both languages, test in both: the language of a prompt changed AI answers more than any change of wording VOCTOS has measured. In our September 2026 Saudi health study, Google AI Mode’s Arabic and English answers to the same question shared 0.149 of their cited domains on average. Repeats a day later shared 0.602.

The score runs from 0 (no shared source) to 1 (identical), and 14 of the 40 question pairs shared nothing. Rewording moved the sources less. A bare keyword against a full sentence gave 0.258, and Gulf against formal Arabic gave 0.472, inside the repeat’s own noise band of 0.465 to 0.733.

What moved Google AI Mode's sources for the same Saudi health questionWhat moved Google AI Mode's sources for the same Saudi health questionMean overlap of cited domains between paired answers (1.0 = identical sources). The Gulf against formal Arabic resultsits inside the repeat's noise band of 0.465 to 0.733.Same question, asked again 24 hourslater (8 pairs)0.602Gulf against formal Arabic (3 pairs)0.472Full sentence against bare keyword (2pairs)0.258Arabic against English (40 pairs)0.149Source: VOCTOS Research, Saudi health questions in Google AI Mode, signed out, September 2026.

Brand lists follow the same pattern. In our October 2026 Egypt appliance study, ChatGPT’s brand lists scored 0.40 between Arabic and English, against 0.54 for the same Egyptian Arabic questions asked twice.

English also cut the made-in-Egypt share of picks from 33.0% to 24.7%, the kind of tilt a 2024 paper on global and local brand bias in large language models describes.

Arabic also starts from a thin web. W3Techs put it at 0.6% of websites with a known content language in August 2026, against 49.5% for English, a figure our review of Arabic AI citation research builds on.

Does the dialect change which brands get named?

On current evidence, dialect moves the brand list about as much as asking twice, but it changes the voice of the answer. In our October 2026 Egypt appliance study, Egyptian against formal Arabic gave ChatGPT brand lists a stability of 0.52, against 0.54 for a plain repeat. Register slips are the bigger risk.

Our 19 October test, described below, shows the slip. Signed-out ChatGPT kept the register of 9 of 10 Arabic prompts, including 3 of 4 Gulf ones; its one miss came back formal.

Google AI Mode kept 5 of 10. None of its 4 Gulf answers was in Gulf Arabic: three came back formal and one switched after a Gulf opening. One Egyptian answer also mixed in formal Arabic.

Did the answer keep the register of the Arabic prompt?Did the answer keep the register of the Arabic prompt?Each 10% is one answer. AI Mode's two mixed answers were a Gulf opening over a formal body and an Egyptian answer thatmixed in formal Arabic.Same register as the promptFormal Arabic for a Gulf promptMixed registerChatGPT, signed out (10Arabic answers)90%10%Google AI Mode (10Arabic answers)50%30%20%Source: VOCTOS test, 19 Oct 2026, run 1: 4 Gulf, 4 Egyptian and 2 formal Arabic prompts per engine.

None of our 8 Gulf answers slipped into Egyptian dialect. That did happen in our October GPT-6 shopping test, on a personalised Plus account used from Cairo. There, 2 of 6 Gulf prompts slipped into Egyptian dialect, one fully and one in its opening phrase. Signed out, the slip we saw ran towards formal Arabic.

So write your core set in the dialect your buyers type, add formal Arabic twins for your top prompts, and log the register of every Arabic answer. These markers tell the registers apart in transliteration; the workbook holds the Arabic spelling.

Meaning Gulf Egyptian Formal Arabic
What’s the best wesh afdal eh ahsan ma afdal
What wesh eh ma
I want abi ayez ureed
Where wein fein ayna
How much kam bikam kam
Now alheen dilwaati alaan
Good zein kwayyis jayyid

The 20-prompt test: two sectors, two engines

We ran 20 library prompts on signed-out ChatGPT and Google AI Mode on 19 October 2026, from 01:16 to 02:25 Cairo time, then once more on each engine. Ten questions covered private dental clinics in Riyadh and online electronics stores in Egypt, each asked in English and eight also in dialect, plus two formal Arabic twins.

Every run went out from Egypt, so every Riyadh prompt named the city. ChatGPT got a new chat per prompt, with search left at its default (it searched on all 20). AI Mode ran through google.com/search?udm=50 with hl set to the prompt’s language and no country parameter. We hashed the prompt file before the first capture.

Group Sector and market Intent Asked in
A1 Private dental clinics, Riyadh Category discovery English, Gulf Arabic, formal Arabic
A2 Same Price (dental implant) English, Gulf Arabic
A3 Same Problem-first (toothache, open now) English, Gulf Arabic
A4 Same Local (female dentist, north Riyadh) English, Gulf Arabic
A5 Same Trust and reviews English
B1 Online electronics stores, Egypt Category discovery (mobile phone) English, Egyptian Arabic, formal Arabic
B2 Same Comparison (laptop, B.TECH or Jumia) English, Egyptian Arabic
B3 Same Trust (Jumia, original products) English, Egyptian Arabic
B4 Same Price (phone in installments, no interest) English, Egyptian Arabic
B5 Same Buyer stage (original iPhone, official warranty) English

One of our analysts coded all 80 answers against written rules. A business offered as a pick counted as recommended, one named only for comparison counted as mentioned, and spellings were merged across Arabic and English. The limits: signed out, two sectors, Riyadh questions sent from Egypt and AI Mode sources read from page data. Both repeats ran in the afternoon rather than at the first run’s hour: ChatGPT from 15:12 to 16:12 that day, AI Mode from 13:41 to 14:13 on 21 October. The AI Mode repeat also used a private Chrome window instead of the first run’s browser, with Google still placing it in Egypt.

What the answers showed

Four results from our first 40 answers are clear even on one run. Every answer came back in the prompt’s language. AI Mode recommended more businesses per answer and never answered a Gulf prompt in Gulf dialect; ChatGPT always cited an outside source. ChatGPT’s sponsored units did not repeat: 13 of 20 answers at night, none of 20 that afternoon.

What 20 answers per engine contained, plus sponsored units in the repeatsWhat 20 answers per engine contained, plus sponsored units in the repeatsAnswers out of 20 per engine, from run 1 except the repeat row. Outside sources exclude Google's own place cards andbusiness profiles in AI Mode.ChatGPT, signed outGoogle AI ModeAnswered in the prompt's languageChatGPT, signed out20 of 20Google AI Mode20 of 20Cited at least one outside sourceChatGPT, signed out20 of 20Google AI Mode16 of 20Showed a sponsored unit, run 1 at night (01:16 to 02:25 Cairo time)ChatGPT, signed out13 of 20Google AI Mode0 of 20Showed a sponsored unit, run 2 in the afternoon (ChatGPT 19 Oct, AI Mode 21 Oct)ChatGPT, signed out0 of 20Google AI Mode0 of 20Ended with a question to the userChatGPT, signed out18 of 20Google AI Mode16 of 20Source: VOCTOS test, 20 prompts per engine from Egypt. Run 1 on both engines, 19 Oct 2026, 01:16 to 02:25 Cairo time;repeats on ChatGPT 19 Oct, 15:12 to 16:12, and on AI Mode 21 Oct, 13:41 to 14:13.

At night, ChatGPT’s sponsored units appeared on 6 of 10 dental answers and 7 of 10 electronics answers. The advertisers were one dental clinic on all six, a phone brand on six and a phone-case seller on one. By the afternoon repeat, signed-out ChatGPT was serving a different, lighter web app; two runs do not show why the ads were gone. Our October 2026 Egypt local study found sponsored units on 35 of 100 ChatGPT answers and none on AI Mode.

Every ChatGPT answer cited at least one outside page, against 16 of 20 on AI Mode. Its other four were Riyadh clinic lists built on Google’s own place cards, which we did not count. AI Mode also cited Facebook in 5 of its 10 Egypt electronics answers and Instagram in 3 dental answers. ChatGPT cited neither.

No answer on either engine refused, gave only general advice or just asked questions back. Most ended by asking the user something: 18 of 20 on ChatGPT and 16 of 20 on AI Mode.

Run 1, 19 Oct 2026 ChatGPT, signed out Google AI Mode
Answers 20 20
In the prompt’s language 20 of 20 20 of 20
Arabic answers in the prompt’s register 9 of 10 5 of 10
Gulf prompts answered in Gulf Arabic 3 of 4 0 of 4
Businesses recommended per answer, mean (median) 2.8 (3) 3.5 (4)
Same, 14 list-style prompts only 4.0 5.0
At least one outside source cited 20 of 20 16 of 20
Sponsored unit shown 13 of 20 (afternoon repeat: 0 of 20) 0 of 20 (repeat: 0 of 20)
Refused, generic or questions only 0 of 20 0 of 20
Ended with a question to the user 18 of 20 16 of 20
Mean answer length 297 words 265 words

Your buyers are asking these questions already. Run these prompts for my brand in Arabic and English

Which businesses got named

In our 19 October test, AI Mode recommended 3.5 businesses per answer against ChatGPT’s 2.8, and 5.0 against 4.0 on the 14 list-style prompts. The evidence behind the names split by sector: Egyptian retailers were cited through their own store pages, Riyadh clinics through directories and Google Maps data.

Our October Egypt local study also found longer AI Mode lists, 6.7 businesses against 3.6. Longer lists give your brand more chances to appear, beside more rivals.

In this test, a store’s own site was cited for 21 of 22 ChatGPT recommendations and 27 of 31 on AI Mode. A clinic’s own site was cited for 1 of 34 and 6 of 39, not counting the Website buttons on AI Mode’s place cards.

How often a recommended business's own website was cited, run 1How often a recommended business's own website was cited, run 1Share of recommendations whose own site was cited in the answer text or source list. Website buttons on AI Mode placecards are not counted.ChatGPT, signed outGoogle AI ModeOnline electronics stores, EgyptChatGPT, signed out95.5%Google AI Mode87.1%Private dental clinics, RiyadhChatGPT, signed out2.9%Google AI Mode15.4%Source: VOCTOS test, 19 Oct 2026, run 1. Shares derived from counts: stores 21 of 22 and 27 of 31, clinics 1 of 34 and6 of 39 recommendations.

For the Riyadh clinics, the evidence behind the picks was directories and map listings, not the clinics’ own websites. One Riyadh dental group was the first pick in 4 of 10 ChatGPT answers and 3 of 10 on AI Mode. In 2 of 10 ChatGPT dental answers, a clinic card named one clinic and linked to another.

For the Riyadh clinics, the evidence behind the picks was directories and map listings, not the clinics' own websites.

The retailers show what your Summary tab will look like. Across six unbranded electronics prompts, B.TECH was recommended in all 6 AI Mode answers and 3 ChatGPT answers, and Jumia in 4 ChatGPT answers but no AI Mode answer. Amazon Egypt was the first pick in 3 of 6 on each engine. Treat one-run gaps like these as leads to recheck.

Engine and language overlaps, read against the noise floor

ChatGPT and AI Mode shared about a quarter of their recommended businesses in run 1 and a fifth in run 2. Each engine shared 0.41 or 0.42 with itself on a repeat, so the engines differ by more than their own noise. Language twins sat below that floor on ChatGPT; on AI Mode the result is mixed.

Comparison (overlap: 0 = nothing shared, 1 = identical) Pairs Brand overlap Source overlap
ChatGPT against AI Mode, same prompt (this test) 14 brand (run 2: 13 source) 0.25 (run 2: 0.20) Not measured (run 2: 0.30)
English against dialect twin, ChatGPT (this test) 5 brand, 8 source 0.27 (run 2: 0.31) 0.35 (run 2: 0.49)
English against dialect twin, AI Mode (this test) 5 brand, 6 source (run 2: 5 source) 0.24 (run 2: 0.37) 0.23 (run 2: 0.44)
English against formal Arabic twin, ChatGPT (this test) 2 0.69 (run 2: 0.21) 0.49 (run 2: 0.29)
English against formal Arabic twin, AI Mode (this test) 2 brand, 1 source 0.50 (run 2: 0.54) 0.25 (run 2: 0.67)
ChatGPT against AI Mode, Egypt local picks (Oct 2026) 100 0.27 Not measured
Same question twice, ChatGPT local picks, Egypt (Oct 2026) 100 0.44 Not measured
Same question twice, ChatGPT brand lists, Egypt appliances (Oct 2026) 60 questions 0.54 (stability score) Not measured
Same question a day later, AI Mode, Saudi health (Sep 2026) 8 Not measured 0.602
Same 20 prompts about 14 hours later, ChatGPT (this test) 14 brand, 20 source 0.41 (median 0.50) 0.55 (median 0.55)
Same 20 prompts about 60 hours later, AI Mode (this test) 14 brand, 13 source 0.42 (median 0.40) 0.38 (median 0.40)

ChatGPT’s floor of 0.41 sits near the 0.44 from our Egypt local study’s repeat, and our engine figure of 0.25 near that study’s 0.27. AI Mode’s floor was 0.42 for brands and 0.38 for cited domains, under the 0.602 of our September 2026 Saudi health repeat on other prompts. The run 2 engine figure of 0.20 compares answers about 46 hours apart, so it mixes engine and time. On 14 prompts in two sectors, the engines overlapped less with each other than with themselves: a measured pattern on a small sample, not a cause.

On ChatGPT, English and dialect twins shared 0.27 of their recommended brands in run 1 and 0.31 in run 2, both below the 0.41 floor. Cited domains followed the same order, at 0.35 and 0.49 against a floor of 0.55. On AI Mode, dialect twins stayed below the 0.42 brand floor (0.24 and 0.37), but on five pairs their cited domains went from 0.23, under the 0.38 floor, to 0.44, above it. The two formal Arabic pairs per engine are too few to read either way: ChatGPT’s brand overlap on them fell from 0.69 to 0.21 between runs.

How to run the prompts yourself

Run the prompts by hand, signed out, in a fresh chat for each one, with the interface language matched to the prompt. Log every answer the same way and repeat the whole set at least 12 hours later. Our 20 prompts on two engines, 40 answers in all, took 69 minutes to capture.

Start from the 100-prompt workbook (XLSX) or the CSV version, then work through six steps.

  1. Fill the placeholders on the How to use tab with your category, city, brand and competitors, spelled the way buyers type them.
  2. Pick 20 prompts: 10 English and 10 Arabic in your buyers’ dialect, mostly unbranded.
  3. In ChatGPT, sign out, open a new chat for every prompt and clear any suggested text first. In our October Egypt studies, signed-out ChatGPT hit its message limit after 74 to 75 questions, so keep each sitting to 40.
  4. In Google AI Mode, load google.com/search?udm=50 with hl=en for English or hl=ar for Arabic. In our September Saudi study, four of seven English prompts came back in Arabic until we set hl=en.
  5. Log every business named, in order, with its sources, any sponsored unit and the answer’s language and dialect.
  6. Repeat the set at least 12 hours later, ideally the next day at the same hour, then read the Summary tab.

How to score each answer

Score each answer from 0 to 5 in the workbook’s Score log tab. You earn one point each for being named, recommended rather than mentioned, among the first three names, the first pick and cited through your own site. The Summary tab turns those rows into rates by engine.

Score log column What you enter Points
Your brand named? Y or N 1 if Y
Recommended or only mentioned? R or M 1 if R
Position of your brand 1 = first name in the answer 1 if in the first three
First pick? Y or N 1 if Y
Your own site cited? Y or N 1 if Y
Answer score Sum of the five 0 to 5
Facts about you correct? Y, N or NA Not scored; any N becomes a correction task
Answer in the language asked? Y or N Not scored; tracks language and dialect slips
Sponsored unit shown? Y or N Not scored; context only

A brand recommended first with its own site cited scores 5; one named fourth, only as a comparison, scores 1. The Summary tab averages the scores and reports visibility, recommendation, first-pick and own-site rates for ChatGPT, Google AI Mode, Gemini and Perplexity.

Three rules keep the totals honest. Merge every spelling of a brand, ignore names that appear only in a closing question or tap-to-answer chip, and keep branded prompts in their own rows. Asked whether to buy a laptop from B.TECH or Jumia, neither engine recommended a third store in any of our 4 answers.

The position point borrows, in simple form, from the position-adjusted measure in the Princeton GEO paper.

How often should you re-run the prompts?

Run every prompt twice before you read anything into a change, then repeat the set monthly, and weekly in a sales season such as Ramadan or White Friday. Vendor studies disagree on repeats. Evertune puts the margin of error at 27 points after 5 runs of one prompt; Meikai found one daily run enough across 500 prompts.

The two measure different things. Evertune’s July 2026 analysis of 10,700 ChatGPT prompts shows how uncertain one prompt’s share is: 27 points at 5 runs, 12 at 12 and about 6 at 100.

Meikai’s 250,000 responses across five engines (13 to 22 July 2026) put a broad portfolio’s citation share at 20.52% with one run a day or ten. Extra runs only cut the error from 0.17 to 0.05 points. Both companies sell measurement.

Sources drift as well. Profound compared the domains cited for about 80,000 prompts per platform in June and July 2025. A month later, 59.3% were new on Google AI Overviews, 54.1% on ChatGPT, 53.4% on Microsoft Copilot and 40.5% on Perplexity. A quarterly check misses most of that movement.

Share of cited domains that were new one month later, same promptsShare of cited domains that were new one month later, same promptsDomains cited in July that had not been cited for the same prompt in June.Google AI Overviews59.3%ChatGPT54.1%Microsoft Copilot53.4%Perplexity40.5%Source: Profound, about 80,000 prompts per platform, 11 to 13 June against 11 to 13 July 2025. Vendor research.

So our rule for Arab-market tests is two runs at least 12 hours apart, then the same prompts monthly with the wording frozen. If you are diagnosing one prompt, run it more often; if you track 50 or 100 prompts, breadth matters more than repeats.

Our read

Test in Arabic and English, and treat a single run as a draft rather than a finding. Language moved AI answers more than any change of wording we have measured. Dialect mostly changed the voice of the answer, and without a repeat you cannot tell a real gap from noise.

Test both engines as well. With both repeats in, ChatGPT and AI Mode shared fewer picks with each other than with themselves, so a brand that shows up on one tells you little about the other.

Group prompts by intent, because wording matters less than the question behind it. In the SparkToro and Gumshoe data, 142 people wrote headphone prompts that barely resembled each other (semantic similarity 0.081). Four brands still appeared in 55% to 77% of the 994 answers, and Peec’s documentation says the same: same-intent prompts converge.

Treat ads as part of the answer now. ChatGPT showed a sponsored unit beside 13 of our 20 signed-out answers at night, so a rival the answer leaves out can still sit beside you as an ad. The afternoon repeat showed none in 20, so log ads on every run, apart from mentions, as the score sheet does. Check facts before you chase mentions: a clinic card that opens the wrong clinic sends a patient to someone else.

Language moved AI answers more than any change of wording we have measured.

FAQ

How many prompts do I need to track my brand in ChatGPT?

Tool makers recommend starting with 20 to 100 prompts per market and language. SE Ranking says 20 to 40, Semrush includes 25 in its base plan and Otterly advises about 100 per country, language and category. In Arab markets count Arabic and English separately, start with 20 prompts and track about 50 each month.

Should I test AI visibility in Arabic or English in Saudi Arabia?

Both. In our September 2026 Saudi health study, Google AI Mode’s Arabic and English answers to the same question shared only 0.149 of their cited domains on average. Repeats a day later shared 0.602. Write the Arabic set in the dialect your customers type, and add formal Arabic twins for your top prompts.

ChatGPT gives me a different answer every time. Is tracking pointless?

No, but one answer means little. In 2,961 runs of 12 prompts, SparkToro and Gumshoe found ChatGPT and Google’s AI returned the same brand list less than once in 100 repeats. Read visibility as a share of many answers, run each prompt at least twice, and compare runs instead of single screenshots.

Does ChatGPT answer Gulf Arabic questions in Gulf dialect?

Mostly, when signed out. In our 19 October 2026 test, signed-out ChatGPT matched the register of 9 of 10 Arabic prompts, including 3 of 4 Gulf ones. Google AI Mode answered all 4 Gulf prompts in formal or mixed Arabic. In an earlier test on a personalised Plus account used from Cairo, 2 of 6 Gulf answers slipped into Egyptian.

Can I test AI visibility for free without a tool?

Yes. Sign out, open a new chat for every prompt, run the same prompts twice at least 12 hours apart and log each answer in a free sheet like ours. Paid tools add scale, daily scheduling and history; they do not change what a single answer can tell you, which is very little on its own.

Start with the questions your buyers actually type

Asking whether AI engines know your brand is too broad to act on. Ask instead which of your buyers’ questions, in which language and dialect, return a list without you, and what those answers cite in your place. The library and score sheet are built to answer that, one market at a time.

If you sell in one market and one language, run the 20-prompt starter set twice this week and correct any wrong facts first. Then work through what to change once you know where you are missing.

If you sell across several Arab markets in both languages, the full library across engines and dialects becomes a monthly job, which VOCTOS can run for you.

Either way, freeze the wording between runs, or you end up measuring your own edits instead of the engines.

We can run all 100 prompts across both engines and every dialect your buyers use. Map my brand’s gaps in Arabic and English AI answers

Everything else we have run on GEO & AEO

Written by whoever ran the work, not a content team

38 articles
All articles
Previous articleNext article
Did you like the article?
Share:

Real Results

More Wins
Delta Medical Labs

SEO & GEO for Delta Medical Labs

GEO
SEO & GEO for Delta Medical Labs

1M+ monthly organic visits, 159,200 ranked keywords, 5,570 pages cited by AI engines, and 1,186 #1 rankings in KSA.

VPNly

Arabic SEO, GEO & guest posting for VPNly

SEO
Arabic SEO, GEO & guest posting for VPNly

65K peak monthly organic visits, 55 guest posts, and 149 pages cited by AI engines.

Ogaei Virtual Care

SEO & GEO for Ogaei Virtual Care

GEO
SEO & GEO for Ogaei Virtual Care

From 3 to 7,931 monthly organic clicks in 7 months, 2.9M+ impressions, and 40+ keywords in Google’s Top 10.

Eduverse

SEO for Eduverse

SEO
SEO for Eduverse

12× search traffic in 10 months, 45% of target keywords in the top 5, and 55% in the top 10.

Read also

All GEO & AEO
  • How SEO and GEO Let a Saudi Medical Lab Cut Google Ads Spend by 87%

  • Which Websites Does Gemini Cite for Arabic Questions? A 2026 Test

  • GEO, AEO and AI Search Glossary for Arab Markets: 70 Terms Explained

  • Google and Bing’s New AI Training Controls: What Arab Publishers Must Set by 1 March 2027